Update server.js via upload script - 2026-10-15 10:20:14
This commit is contained in:
parent
53370386e2
commit
4ef9398e47
38
server.js
38
server.js
|
|
@ -108,8 +108,16 @@ function saveAiConfig(apiKey) {
|
|||
}
|
||||
|
||||
async function callDeepSeek(apiKey, messages, opts) {
|
||||
const { maxTokens = 600, temperature = 0.7, model = 'deepseek-v4-flash' } = opts || {};
|
||||
const resp = await fetch(`${DEEPSEEK_API_BASE}/chat/completions`, {
|
||||
const { maxTokens = 600, temperature = 0.7, model = 'deepseek-v4-flash', timeoutMs = 120000 } = opts || {};
|
||||
// 超时保护:没有这个的话,DeepSeek 一旦卡住或输出失控(一直生成到 max_tokens 上限),请求会无限期挂着,
|
||||
// 前端就一直转圈。超时后直接报错,用户可以重试。
|
||||
const controller = new AbortController();
|
||||
const timer = setTimeout(() => controller.abort(), timeoutMs);
|
||||
const startedAt = Date.now();
|
||||
let resp;
|
||||
try {
|
||||
resp = await fetch(`${DEEPSEEK_API_BASE}/chat/completions`, {
|
||||
signal: controller.signal,
|
||||
method: 'POST',
|
||||
headers: {
|
||||
'Content-Type': 'application/json',
|
||||
|
|
@ -126,12 +134,25 @@ async function callDeepSeek(apiKey, messages, opts) {
|
|||
thinking: { type: 'disabled' },
|
||||
}),
|
||||
});
|
||||
} catch (e) {
|
||||
clearTimeout(timer);
|
||||
if (e.name === 'AbortError') throw new Error(`DeepSeek 超过 ${Math.round(timeoutMs / 1000)} 秒没有返回结果,请重试`);
|
||||
throw e;
|
||||
}
|
||||
if (!resp.ok) {
|
||||
clearTimeout(timer);
|
||||
const errText = await resp.text().catch(() => '');
|
||||
throw new Error(`DeepSeek API error ${resp.status}: ${errText.slice(0, 200)}`);
|
||||
}
|
||||
const data = await resp.json();
|
||||
let data;
|
||||
try { data = await resp.json(); }
|
||||
catch (e) {
|
||||
if (e.name === 'AbortError') throw new Error(`DeepSeek 超过 ${Math.round(timeoutMs / 1000)} 秒没有返回结果,请重试`);
|
||||
throw e;
|
||||
} finally { clearTimeout(timer); }
|
||||
const choice = data.choices && data.choices[0];
|
||||
// 记录耗时和 token 用量,以后再遇到"很慢"可以直接在日志里看到原因
|
||||
console.log(`[DeepSeek] ${Math.round((Date.now() - startedAt) / 1000)}s, finish=${choice && choice.finish_reason}, usage=${JSON.stringify(data.usage || {})}`);
|
||||
const content = choice && choice.message && choice.message.content;
|
||||
if (!content) {
|
||||
const reason = choice && choice.finish_reason ? `(finish_reason: ${choice.finish_reason})` : '';
|
||||
|
|
@ -602,6 +623,9 @@ moduleDimensions/moduleWeight这组字段里就行,cabinetXxx那组(cabinetM
|
|||
"notes": "其他有用但上面字段装不下的信息,一两句话,没有就填空字符串"
|
||||
}
|
||||
|
||||
【输出精简】上面的 schema 只是字段说明:每条产品记录里,值为空字符串或空数组的字段一律省略不写,只输出原文里真正有值的字段(name 必须输出)。
|
||||
这样输出更短、更快,也不会因为输出太长被截断。
|
||||
|
||||
只返回一个JSON数组,不要有任何解释文字、不要用markdown代码块包裹、不要省略任何一个识别到的型号。`;
|
||||
|
||||
app.post('/api/ai/parse-catalog', requireAuth, aiLimiter, (req, res) => {
|
||||
|
|
@ -670,7 +694,13 @@ app.post('/api/ai/parse-catalog', requireAuth, aiLimiter, (req, res) => {
|
|||
const content = await callDeepSeekWithRetry(cfg.apiKey, [
|
||||
{ role: 'system', content: CATALOG_PARSE_SYSTEM_PROMPT },
|
||||
{ role: 'user', content: userContent },
|
||||
], { maxTokens: 64000, temperature: 0.2 });
|
||||
], {
|
||||
// 按文档长度给输出预算:短文档(比如一页HTML规格表)不需要64000这么大的上限,
|
||||
// 上限太大时一旦模型输出失控,会一直生成十几分钟才结束
|
||||
maxTokens: Math.min(64000, Math.max(12000, Math.round(clippedText.length * 1.5))),
|
||||
temperature: 0.2,
|
||||
timeoutMs: 300000, // 5分钟
|
||||
});
|
||||
const { data: products, truncated } = extractJson(content);
|
||||
if (!Array.isArray(products)) throw new Error('AI没有返回预期的数组格式,请重试');
|
||||
// 双保险:即使AI没完全听话,这里直接强制覆盖,保证和用户预选的一致
|
||||
|
|
|
|||
Loading…
Reference in New Issue